A Python library for creating super fancy Unix daemons
daemonocle is a library for creating your own Unix-style daemons written in Python. It solves many problems that other daemon libraries have and provides some really useful features you don't often see in other daemons.
Table of Contents
To install via pip:
pip install daemonocle
Or download the source code and install manually:
git clone https://github.com/jnrbsn/daemonocle.git cd daemonocle/ python setup.py install
Here's a really really basic example:
import sys
import time
import daemonocle
# This is your daemon. It sleeps, and then sleeps again.
def main():
while True:
time.sleep(10)
if __name__ == '__main__':
daemon = daemonocle.Daemon(
worker=main,
pid_file='/var/run/daemonocle_example.pid',
)
daemon.do_action(sys.argv[1])
And here's the same example with logging and a Shutdown Callback:
import logging
import sys
import time
import daemonocle
def cb_shutdown(message, code):
logging.info('Daemon is stopping')
logging.debug(message)
def main():
logging.basicConfig(
filename='/var/log/daemonocle_example.log',
level=logging.DEBUG, format='%(asctime)s [%(levelname)s] %(message)s',
)
logging.info('Daemon is starting')
while True:
logging.debug('Still running')
time.sleep(10)
if __name__ == '__main__':
daemon = daemonocle.Daemon(
worker=main,
shutdown_callback=cb_shutdown,
pid_file='/var/run/daemonocle_example.pid',
)
daemon.do_action(sys.argv[1])
And here's what it looks like when you run it:
user@host:~$ python example.py start Starting example.py ... OK user@host:~$ python example.py status example.py -- pid: 1234, status: running, uptime: 1m, %cpu: 0.0, %mem: 0.0 user@host:~$ python example.py stop Stopping example.py ... OK user@host:~$ cat /var/log/daemonocle_example.log 2014-05-04 12:39:21,090 [INFO] Daemon is starting 2014-05-04 12:39:21,091 [DEBUG] Still running 2014-05-04 12:39:31,091 [DEBUG] Still running 2014-05-04 12:39:41,091 [DEBUG] Still running 2014-05-04 12:39:51,093 [DEBUG] Still running 2014-05-04 12:40:01,094 [DEBUG] Still running 2014-05-04 12:40:07,113 [INFO] Daemon is stopping 2014-05-04 12:40:07,114 [DEBUG] Terminated by SIGTERM (15)
For more details, see the Detailed Usage section below.
If you think about it, a lot of Unix daemons don't really know what the hell they're doing. Have you ever found yourself in a situation that looked something like this?
user@host:~$ sudo example start starting example ... ok user@host:~$ ps aux | grep example user 1234 0.0 0.0 1234 1234 pts/1 S+ 12:34 0:00 grep example user@host:~$ sudo example start starting example ... ok user@host:~$ echo $? 0 user@host:~$ tail -f /var/log/example.log ...
Or something like this?
user@host:~$ sudo example stop stopping example ... ok user@host:~$ ps aux | grep example user 123 0.0 0.0 1234 1234 ? Ss 00:00 0:00 /usr/local/bin/example user 1234 0.0 0.0 1234 1234 pts/1 S+ 12:34 0:00 grep example user@host:~$ sudo example stop stopping example ... ok user@host:~$ ps aux | grep example user 123 0.0 0.0 1234 1234 ? Ss 00:00 0:00 /usr/local/bin/example user 1240 0.0 0.0 1234 1234 pts/1 S+ 12:34 0:00 grep example user@host:~$ sudo kill -9 123 ...
Or something like this?
user@host:~$ sudo example status Usage: example {start|stop|restart} user@host:~$ ps aux | grep example ...
These are just a few examples of unnecessarily common problems. It doesn't have to be this way.
Note: You might be thinking, "Why not just write a smarter start/stop shell script wrapper for your daemon that checks whether or not it actually started, actually stopped, etc.?" Seriously? It doesn't have to be this way. I believe daemons should be more self-aware. They should handle their own problems most of the time, and your start/stop script should only be a very thin wrapper around your daemon or simply a symlink to your daemon.
If you've ever dug deep into the nitty-gritty details of how daemonization works, you're probably familiar with the standard "double fork" paradigm first introduced by W. Richard Stevens in the book Advanced Programming in the UNIX Environment. One of the problems with the standard way to implement this is that if the final child dies immediately when it gets around to doing real work, the original parent process (the one that actually had control of your terminal) is long gone. So all you know is that the process got forked, but you have no idea if it actually kept running for more than a fraction of a second. And let's face it, one of the most likely times for a daemon to die is immediately after it starts (due to bad configuration, permissions, etc.).
The next problem mentioned in the section above is when you try to stop a daemon, it doesn't
actually stop, and you have no idea that it didn't actually stop. This happens when a process
doesn't respond properly to a SIGTERM
signal. It happens more often than it should. The problem
is not necessarily the fact that it didn't stop. It's the fact that you didn't know that it didn't
stop. The start/stop script knows that it successfully sent the signal and so it assumes success.
This also becomes a problem when your restart
command blindly calls stop
and then start
,
because it will try to start a new instance of the daemon before the previous one has exited.
These are the biggest problems most daemons have in my opinion. daemonocle solves these problems and provides many other "fancy" features.
The problem with the daemon immediately dying on startup and you not knowing about it is solved by
having the first child (the immediate parent of the final child) sleep for one second and then call
os.waitpid(pid, os.WNOHANG)
to see if the process is still running. This is what daemonocle
does. So if you're daemon dies within one second of starting, you'll know about it.
This problem with the daemon not stopping and you not knowing about it is solved by simply waiting
for the process to finish (with a timeout). This is what daemonocle does. (Note: When a timeout
occurs, it doesn't try to send a SIGKILL
. This is not always what you'd want and often not a
good idea.)
Below are some other useful features that daemononcle provides that you might not find elsewhere.
There is a status
action that not only displays whether or not the daemon is running and its
PID, but also the uptime of the daemon and the % CPU and % memory usage of all the processes in the
same process group as the daemon (which are probably its children). So if you have a daemon that
launches mulitple worker processes, the status
action will show the % CPU and % memory usage of
all the workers combined.
It might look something like this:
user@host:~$ python example.py status example.py -- pid: 1234, status: running, uptime: 12d 3h 4m, %cpu: 12.4, %mem: 4.5
You can even get JSON output if you call the action like this:
daemon.do_action('status', json=True)
If you use the Integration with click described below, this option is available via the
--json
CLI option. You can also just get a dict
directly and programatically without
printing it to STDOUT by calling Daemon.get_status()
.
Have you ever tried to restart a daemon only to realize that it's not actually running? Let me
guess: it just gave you an error and didn't start the daemon. A lot of the time this is not a
problem, but if you're trying to restart the daemon in an automated way, it's more annoying to have
to check if it's running and do either a start
or restart
accordingly. With daemonocle, if
you try to restart a daemon that's not running, it will give you a warning saying that it wasn't
running and then start the daemon. This is often what people expect.
Daemons that use daemonocle have the ability to reload themselves by simply calling
daemon.reload()
where daemon
is your daemonocle.Daemon
instance. The execution of the
current daemon halts wherever daemon.reload()
was called, and a new daemon is started up to
replace the current one. From your code's perspective, it's pretty much the same as a doing a
restart
except that it's initiated from within the daemon itself and there's no signal handling
involved. Here's a basic example of a daemon that watches a config file and reloads itself when the
config file changes:
import os
import sys
import time
import daemonocle
class FileWatcher(object):
def __init__(self, filename, daemon):
self._filename = filename
self._daemon = daemon
self._file_mtime = os.stat(self._filename).st_mtime
def file_has_changed(self):
current_mtime = os.stat(self._filename).st_mtime
if current_mtime != self._file_mtime:
self._file_mtime = current_mtime
return True
return False
def watch(self):
while True:
if self.file_has_changed():
self._daemon.reload()
time.sleep(1)
if __name__ == '__main__':
daemon = daemonocle.Daemon(pid_file='/var/run/daemonocle_example.pid')
fw = FileWatcher(filename='/etc/daemonocle_example.conf', daemon=daemon)
daemon.worker = fw.watch
daemon.do_action(sys.argv[1])
You may have noticed from the Basic Usage section above that a shutdown_callback
was defined.
This function gets called whenever the daemon is shutting down in a catchable way, which should be
most of the time except for a SIGKILL
or if your server crashes unexpectedly or loses power or
something like that. This function can be used for doing any sort of cleanup that your daemon needs
to do. Also, if you want to log (to the logger of your choice) the reason for the shutdown and the
intended exit code, you can use the message
and code
arguments that will be passed to your
callback (your callback must take these two arguments).
This is not particularly interesting per se, but it's worth noting that in non-detached mode, your
daemon will do everything else you've configured it to do (i.e. setuid
, setgid
, chroot
,
etc.) except actually detaching from your terminal. So while you're testing, you can get an
extremely accurate view of how your daemon will behave in the wild. It's also worth noting that
self-reloading works in non-detached mode, which was a little tricky to figure out initially.
One of the things that daemons typically do is close all open file descriptors and establish new
ones for STDIN
, STDOUT
, STDERR
that just point to /dev/null
. This is fine most of
the time, but if your worker is an instance method of a class that opens files in its __init__()
method, then you'll run into problems if you're not careful. This is also a problem if you're
importing a module that leaves open files behind. For example, importing the
random standard library module in Python 3
results in an open file descriptor for /dev/urandom
.
Since this "feature" of daemons often causes more problems than it solves, and the problems it
causes sometimes have strange side-effects that make it very difficult to troubleshoot, this feature
is optional and disabled by default in daemonocle via the close_open_files
option.
The daemonocle.Daemon
class is the main class for creating a daemon using daemonocle. Here's the
constructor signature for the class:
class daemonocle.Daemon(
name=None, worker=None, detach=True,
pid_file=None, work_dir='/', stdout_file=None, stderr_file=None, chroot_dir=None,
uid=None, gid=None, umask=0o22, close_open_files=False,
shutdown_callback=None, stop_timeout=10)
And here are descriptions of all the arguments:
name
- The name of your program to use in output messages. Default:
os.path.basename(sys.argv[0])
worker
- The function that does all the work for your daemon.
detach
- Whether or not to detach from the terminal and go into the background. See Non-Detached Mode
for more details. Default:
True
pid_file
- The path to a PID file to use. It's not required to use a PID file, but if you don't, you won't be able to use all the features you might expect. Make sure the user your daemon is running as has permission to write to the directory this file is in.
work_dir
- The path to a directory to change to when the daemon starts. Note that a file system cannot be
unmounted if a process has its working directory on that file system. So if you change the
default, be careful about what you change it to. Default:
"/"
stdout_file
If provided when
detach=True
, the STDOUT stream will be redirected (appended) to the file at the given path. In non-detached mode, this argument is ignored.New in version 1.1.0.
stderr_file
If provided when
detach=True
, the STDERR stream will be redirected (appended) to the file at the given path. In non-detached mode, this argument is ignored.New in version 1.1.0.
chroot_dir
- The path to a directory to set as the effective root directory when the daemon starts. The default is not to do anything.
uid
- The user ID to switch to when the daemon starts. The default is to not switch users.
gid
- The group ID to switch to when the daemon starts. The default is to not switch groups.
umask
- The file creation mask ("umask") for the process. Default:
0o022
close_open_files
- Whether or not to close all open files when the daemon detaches. Default:
False
shutdown_callback
- This will get called anytime the daemon is shutting down. It should take a
message
and acode
argument. The message is a human readable message that explains why the daemon is shutting down. It might useful to log this message. The code is the exit code with which it intends to exit. See Shutdown Callback for more details. stop_timeout
- Number of seconds to wait for the daemon to stop before throwing an error. Default:
10
The default actions are start
, stop
, restart
, and status
. You can get a list of
available actions using the daemonocle.Daemon.list_actions()
method. The recommended way to call
an action is using the daemonocle.Daemon.do_action(action)
method. The string name of an action
is the same as the method name except with dashes in place of underscores.
If you want to create your own actions, simply subclass daemonocle.Daemon
and add the
@daemonocle.expose_action
decorator to your action method, and that's it.
Here's an example:
import daemonocle
class MyDaemon(daemonocle.Daemon):
@daemonocle.expose_action
def full_status(self):
"""Get more detailed status of the daemon."""
pass
Then, if you did the basic daemon.do_action(sys.argv[1])
like in all the examples above, you can
call your action with a command like python example.py full-status
.
daemonocle also provides an integration with click, the "Command Line
Interface Creation Kit". The integration is in the form of a custom command class
daemonocle.cli.DaemonCLI
that you can use in conjunction with the @click.command()
decorator
to automatically generate a command line interface with subcommands for all your actions. It also
automatically daemonizes the decorated function. The decorated function becomes the worker, and the
actions are automatically mapped from click to daemonocle.
Here's an example:
import time
import click
from daemonocle.cli import DaemonCLI
@click.command(cls=DaemonCLI, daemon_params={'pid_file': '/var/run/example.pid'})
def main():
"""This is my awesome daemon. It pretends to do work in the background."""
while True:
time.sleep(10)
if __name__ == '__main__':
main()
Here are all the help pages for the default actions:
user@host:~$ python example.py --help Usage: example.py [OPTIONS] COMMAND [ARGS]... This is my awesome daemon. It pretends to do work in the background. Options: --help Show this message and exit. Commands: start Start the daemon. stop Stop the daemon. restart Stop then start the daemon. status Get the status of the daemon. user@host:~$ python example.py start --help Usage: example.py start [OPTIONS] Start the daemon. Options: --debug Do NOT detach and run in the background. --help Show this message and exit. user@host:~$ python example.py stop --help Usage: example.py stop [OPTIONS] Stop the daemon. Options: --timeout INTEGER Number of seconds to wait for the daemon to stop. Overrides "stop_timeout" from daemon definition. --force Kill the daemon uncleanly if the timeout is reached. --help Show this message and exit. user@host:~$ python example.py restart --help Usage: example.py restart [OPTIONS] Stop then start the daemon. Options: --timeout INTEGER Number of seconds to wait for the daemon to stop. Overrides "stop_timeout" from daemon definition. --force Kill the daemon forcefully after the timeout. --debug Do NOT detach and run in the background. --help Show this message and exit. user@host:~$ python example.py status --help Usage: example.py status [OPTIONS] Get the status of the daemon. Options: --json Show the status in JSON format. --fields TEXT Comma-separated list of process info fields to display. --help Show this message and exit.
The daemonocle.cli.DaemonCLI
class also accepts a daemon_class
argument that can be a
subclass of daemonocle.Daemon
. It will use your custom class, automatically create subcommands
for any custom actions you've defined, and use the docstrings of the action methods as the help text
just like click usually does.
This integration is entirely optional. daemonocle doesn't enforce any sort of argument parsing. You
can use argparse, optparse, or just plain sys.argv
if you want.
Starting with version 1.1.0, you can also use a couple different shorter ways of invoking the CLI.
Like this:
from daemonocle.cli import cli
@cli(pid_file='/var/run/example.pid')
def main():
"""Do stuff"""
...
if __name__ == '__main__':
main()
Or like this:
from daemonocle import Daemon
def main():
"""Do stuff"""
...
if __name__ == '__main__':
daemon = Daemon(worker=main, pid_file='/var/run/example.pid')
daemon.cli()
The above two examples are equivalent. Use whichever way works best for you.
Please create an issue on GitHub.